DNA Models Enable Sequence Reconstruction Via Embeddings
AI-assisted, source-derived brief produced by the Let's Data Science Automated News Desk. The source material used is linked on this page.
- Source event:
- first reported
- LDS brief:
- publication time is not available in the public LDS lifecycle record
A March 6, 2026 arXiv preprint by Sofiane Ouaari et al. evaluates reconstruction risks from embeddings produced by DNA foundation models. Testing DNABERT-2, Evo 2, and Nucleotide Transformer v2, authors find per-token embeddings enable near-perfect sequence reconstruction, while mean-pooled embeddings degrade with length yet remain above random baselines. Results highlight urgent privacy risks for Embeddings-as-a-Service in genomics.
Key Points
- 1Demonstrate near-perfect sequence reconstruction from per-token embeddings across three DNA foundation models
- 2Show that embedding–sequence similarity predicts reconstruction success, producing >90% similarity for short sequences
- 3Warn that sharing embeddings via EaaS risks genomic privacy; adopt privacy-aware model designs and controls
Scoring Rationale
Strong methodological novelty and practical attack demonstrations with released code; limited by preprint status and pending peer review.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems